===== (round 2) ===== Reviewer #1: Thank you for addressing my comments. The revisions are promising but I would like to encourage you to pursue one suggestion. Yes, you've linked the DHS survey but it will still be nice to describe the kinds of variables that go in briefly - say in just a couple of sentences. One, to make the point that you have a valid reason to consider all 6500 variables as they may relate to IUD in some way (I know this is a hypothesis-free approach but you'll agree that you'd not want readers to take away the message that this exercise is just a brute data mining effort). And two, it's only going to help enhance the reach and relevance of your paper to non-DHS/FP-focused health and other readers as well who might be interested in your approach. Otherwise, everything else is satisfactory. All the best. Reviewer #3: The authors have considerably strengthened the paper this revision through better contextualizing decisions and positioning this work in the greater scheme of IUD research and policy within India. I believe this article is ready for submission given the significant changes from the original submitted manuscript. ===== (round 1) ===== Reviewer #1: The paper is a great contribution to the literature around rigorously understanding what drives uptake an important spacing method, which despite being available for such a long time is barely used. I particularly appreciate this paper's contribution in introducing and applying some sophisticated and advanced ML techniques, which is not very common. The insights from the paper are also very relevant and actionable. In general, I suggest that authors use shorter sentences throughout the manuscript. Long sentences can make it hard for the readers to follow and comprehend with ease. Also, the grammar can be improved in some places. I have provided additional suggestions on how the paper can be further improved. I hope that the authors find it useful. Line 10: Suggesting changing the word "power" to something else. Power can confuse readers with statistical power unless the authors meant so. Line 23: Can add male engagement, couple dynamics to foster the attention of readers Line 46: Client satisfaction is not supply side. Lines 52-55: Critical demand-side barriers such as perceived side-effects from IUCD (abdominal pain, bleeding etc.) have not been mentioned. There is a lot of literature available on this. Also, side effects is a big component of the theme that this paper finds as important - FP counselling and FP services. Line 90 and 102: While ML models have the ability to conduct unsupervised analysis with a wide range of variables, close to 6500 variables can seem intimidating. The authors should discuss the pros and cons of both hypothesis-free and hypothesis/framework driven approaches. Also, please describe what are the different variables as 6500 variables is a big number. So, it makes one immediately wonder what all has gone into the models. This information can be presented as supplementary material. Additionally, couldn't there be other redundant variables such as women/man is already sterilized or woman is infertile? These variables will have a direct 1:1 relationship with not using IUCD. Line 104 to 110: The authors provide a very detailed description of each of the ML models later on which is very useful. Since ML models might not be well known to everyone, can the authors upfront briefly describe in relatively simpler terms, as they did for the NN model, why were the other two models used. Example - Lasso was used to reduce the high number of variables. Line 218: add vs data for non-users. Also, are these differences statistically different? Table 1: Add statistical significance using any test of independence measure Line 227: fertility is repeated Table 2: using contraceptives for limiting is a perfect predictor for not using IUCD. I have suggested above to consider it as a redundant variable. The same comment for has never used female sterilization. Or can rationalize why it is okay to include perfect predictors in the ML models you have used because it is not a good practice to include such variables in say, standard statistical regression models. Line 262-264: suggest adding some examples from other work/recommendations on how to increase male engagement and couple dynamics. While this is incredibly important, programs have struggled to achieve this. But there is evidence/examples of how might this be made possible. Since this is the main finding of the study, providing more actionable information to reader can be very valuable. Line277-280: the sentence is long and unclear. Lines 297-298: can you add a citation to support the sentence. Lines 303-312: not sure how is this a limitation. Can be appropriately added to the discussion section. Reviewer #2: Thank you for giving me the opportunity to review this manuscript. The manuscript presents a novel way of interpreting data from India's National Family Health Survey concerning uptake and use of IUDs. I am primarily able to offer a response from the perspectives of previous qualitative analyses and to ask whether your method can capture any of these dynamics, or whether this is a limitation with the method or with the dataset used in the analysis. The categories that emerge from your analysis as most important in use of IUD would seem to be those most rooted in decision-making processes at a couple level in the present moment or recent past (i.e., as cited in the abstract, shared family planning (FP), receipt of FP services and desire for no more children, higher wealth and education, and receipt of maternal and child health services). Where the analysis seems to have less to offer lies more in the realm of history, religion, and culture. For example, can this method account for knowledge of risks associated with past versions of the IUD (though there are hints of this in the fears of infertility cited line 272-274) or with the history of IUDs inserted without consent in medical facilities, often at the time of birth? Does the model offer perspectives on how the status of an IUD as a primarily provider-controlled contraceptive method and one that lies between reversible and non-reversible due to its long efficacy and need for professional removal, may accord with religious sentiments about fertility? Can this method of analysis show how ideas about bodily integrity, health, and perceptions of how contraception may interfere with intercourse (fears of infection, of the device moving during intercourse, or of harming a partner) might impact the likelihood of adopting an IUD? Perhaps the questions I have posed here are an indication that the manuscript would benefit from saying more about the limitations of working with the NFHS as source material and about how insight gained from machine learning methods could and ought to be combined with insight provided by the use of other methods. Reviewer #3: This paper offers insight into determinants of IUDS use in India, specifically highlighting the receipt of family planning services, desires for no more children, higher wealth, and education as factors influencing IUD uptake. This is an interesting and original paper on an important topic. However, I think that there are some methodological and analytic questions that must be addressed. I think the authors need to acknowledge the limitations of their data, and what they can and cannot say with it. My main concern is the assumptions underlying the methodological analysis. Both Lasso and Ridge Regression assume that the variables have linear relationship with the target which is not sufficiently established in this article. Additionally, there is no analysis to suggest that multicollinearity does not exist amongst the variables utilised in the study. While your approach is interesting, a statistically rigorous explanation needs to accompany the thematic analysis and the interpretation of data. Similarly, from lines 207-214, the authors describe identifying a group of variables of a theme when the number of variables within a given theme was at least 5% of the total number of identified variables above the knee point of the coefficient curve. However, the selection of 5% is not rationalized or a generally agreed upon threshold. Specifically, it is not clear what exactly is the 'adequate number of coherent themes' mentioned on line 210. There is a need for more details on the variables included within the article. In the Measures section, the authors note that they 6,500 variables after categorizing and converting variables. However, this number needs to be contextualized to detail the range of variables and the type of variables included within the sample. Moreover, the variables referenced in the findings need to be explained in greater detail. You note that you removed irrelevant and redundant variables but there should be a more thorough review of the relevant variables used. For example, variables such as 'knows about traditional method of contraception' leaves us with questions about what is considered a 'traditional' method of birth control. I offer other suggestions of how the paper might be strengthened below: - There is a need for more context regarding IUD services in India and efforts from the Indian government regarding Reproductive, Maternal, Newborn, Child and Adolescent Health (RMNCH+A) Strategy. There is a lack of existing research cited that discuss reasons that may influence adoption and continuation of modern contraceptives. Singal et al. (2022) noted reasons for IUD removal such as side effects or other problems, spouse or family opposition, desire to conceive, decision to undergo sterilisation, and method failure. - Moreover, this article focuses on the low rates of IUD usage but fails to mention all the reasons that IUD rates should be higher, such as how discontinuation rates of IUDS being lower than all other contraceptive methods (see IIPS and ICF 2017). There is a lot of information in the NFHS-5 Report that makes the puzzle of low IUD uptake more interesting which sets the article up well. - This paper is aimed as a non-machine learning audience, which is beneficial considering the readership of SSM. However, despite a clear description of machine learning and neural network approaches, critical concepts such as overfitting and bias should be better explained in the context of the paper. - The data includes 699,686 women aged 15-49 years old. You note that 499,627 were married and asked about contraceptive use. While I understand the focus on married women, I believe you should include a rationale for this sample. Were only married women asked about contraceptive use? - There are inconsistencies between your points of comparison. You often compare IUD user and non-users but then at times focus on only IUD users, such as when you discuss wealth and urbanicity. - Table 1 includes keywords and acronyms (such as 'SC/ST' and 'OBC') which are not discussed in the data section. There is a need for better detailing of this data to enhance reader comprehension upon results section. Your table could also be clearer by distinguishing between when the value is a percentage of the population versus the mean value. - There is a lot of editorializing of what models did and did not work. However, the results should be better formatted to explain the various themes and the grouping opposed to data/regression specific decisions. - The limitation sections acknowledge the choice of machine learning models but ignores some of the key limitations of machine learning as a model for this study and quantitative data more generally. The limitations section acknowledges that the findings reflect themes from variables that account for the most variance but then goes off track by focusing on disproportionate use of female sterilization and the implications of this trend which is not a limitation of the data but a cultural context to consider. Minor Issues: - Line 145: Satopaa et al. without relevant year - Line 198: rural/urban is noted twice